Low-rank quaternion approximation for color image processing.

Area of Science:

• Computer Vision
• Image Processing
• Matrix Approximation

Background:

• Low-rank matrix approximation (LRMA) is successful for grayscale images.
• Current LRMA for color images treats channels independently or concatenated, missing cross-channel correlations.
• Existing methods may not fully exploit RGB channel correlations.

Purpose of the Study:

• Propose a novel low-rank quaternion approximation (LRQA) model for color images.
• Improve color image processing by leveraging cross-channel correlations.
• Enhance performance in tasks like denoising and inpainting.

Main Methods:

• Encode color images as pure quaternion matrices to exploit cross-channel correlations.
• Impose low-rank constraints on the constructed quaternion matrices.
• Develop a general LRQA model using nonconvex functions for singular value estimation.

Main Results:

• LRQA effectively utilizes correlations among RGB channels.
• The proposed LRQA model achieves superior performance in color image denoising.
• LRQA demonstrates better results in color image inpainting compared to state-of-the-art methods.
• Evaluations show improvements in both quantitative metrics and visual quality.

Conclusions:

• The proposed LRQA model offers a significant advancement for color image processing.
• LRQA effectively captures and utilizes inter-channel correlations, outperforming traditional LRMA and sparse representation methods.
• LRQA provides a robust framework for various image restoration tasks.